Context window
The maximum amount of text, measured in tokens, a model can consider at once, covering both your input and its output.
The context window is how much a model can “see” at one time, measured in tokens (roughly, chunks of a few characters). A 128k context window means the model can work with around 128,000 tokens of input and output combined.
A longer context lets a model handle longer documents and conversations, but it is not free: the memory used to hold the context (the KV cache) grows with it, so long contexts consume more VRAM. A model that fits comfortably at short context can run out of memory once the conversation grows long.
Stated maximum context is also not the same as reliable context. Many models handle information less well towards the far end of their window than the headline number suggests.